Opus 5 solves the largest security gap of AI agents

Claude Opus 5 + Auto Mode had 0% successful prompt injection attacks with 129 tests on browser agents. Claude Opus 5 is 3.7% vulnerable without Auto Mode — compared to 2.8% (Fable 5) and 2.6% (Mythos 5) from the top competitors on the Gray Swan IPI Benchmark.

Auto Mode provides two layers of defense: blocking hidden code before the model runs and stopping dangerous code at execution. A successful prompt injection requires an attacker to pass through both layers, resulting in the reported 0% success rate.

Prompt injection is when an attacker injects hidden code into the webpage an AI agent reads. The agent will run that code instead of completing the intended task. OpenAI has said, “this is probably never going to be fully solved.” Anthropic is here to prove otherwise.